Paper
Publication
Scalable Deep Reinforcement Learning Algorithms for Mean Field Games
Abstract

Fictitious Play or online mirror descent algorithms have revealed very efficient to approximate Nash equilibria in multi-agent systems with many agents, whenver the system fits in the framework of the s-called mean field games. We study how these approaches combine efficiently with deep learning architectures and allow to scale the complexity of the considered game dynamics and environments.

ICML2022

Authors' notes